{"id":"54c36e5d-b3b0-4275-a1a5-d54d75d355cd","arxiv_id":"2512.17678","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"YOTO jointly learns a discrete gene subset and a multi-task predictor in one differentiable model, outperforming baselines only at moderate-to-large panel sizes.","lead":"The paper introduces YOTO, a single neural network that picks a small set of genes and predicts cell labels at the same time, so selection and prediction are trained together. A smart generalist might read it to see whether end-to-end, sparse gene selection can beat classical marker-gene pipelines for single-cell data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's own tables contradict 'consistently outperforms': YOTO loses AUROC by >17 pts at k=64 and trails baselines at k=16; prose mischaracterizes these as 'slightly lower' and 'only exception.'","rationale":"The reader's verdict is CONDITIONAL, and my stress-test supports that verdict but for a different central reason. The reader's stated weakest_assumption focuses on the straight-through estimator (STE) in Section 3.1. That is a legitimate theoretical concern, but the empirical evidence from the paper's own tables is more directly load-bearing. The abstract claims 'consistently outperforms state-of-the-art baselines,' yet Tables 5–8 show YOTO losing on AUROC by large margins at k=64 and trailing baselines at k=16 and k=32 on F1. The paper's prose minimizes these losses as 'slightly lower' or 'only exception,' which is an accurate representation issue. If the central claim were merely about the STE, a synthetic experiment might show whether the gradient proxy works. But the paper's own reported numbers already undermine the headline claim, so the most crucial concern is that the evidence as presented does not support the stated superiority. This is consistent with the reader's rationale, which also flags the overclaim and AUROC discrepancy, though the reader chose to emphasize the STE as the weakest assumption. Therefore, I partially agree with the reader: the empirical contradiction is the true load-bearing issue, and it is concrete and checkable. The verdict remains CONDITIONAL: the method may have merit, but the claims must be revised and the results made reproducible before acceptance.","tokens_in":13527,"tokens_out":7140,"duration_ms":73390,"concrete_test":"Reproduce the COVID-PBMC celltype5 experiment at k=64 following the exact preprocessing and training protocol described in Section 4.3 (three seeds, 1000 epochs, temperature annealing) and compute the AUROC. If the reported 80.9±0.2 is reproduced, then YOTO's AUROC is far below the baselines and the 'consistently outperforms' claim is false; if the AUROC differs substantially, the paper's evaluation is non-reproducible and the reported numbers cannot be trusted. This single check settles whether the concern about the headline claim is valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that YOTO 'consistently outperforms state-of-the-art baselines' is contradicted by the paper's own reported results. Table 8 (COVID-PBMC, k=64) shows YOTO's AUROC = 80.9±0.2 versus mRMR-f = 98.3±0.2 and PERSIST = 95.8±0.1 — a 17-point gap that Section 4.4.3 describes as 'slightly lower.' Table 7 (VISp, k=64) shows the same pattern: YOTO AUROC = 91.9±0.1 versus Block HSIC Lasso and mRMR-rf at 99.9±0.0. At small panel sizes, YOTO loses on F1: Table 5 (VISp, k=16) has YOTO at 64.0±4.0 versus PERSIST 74.3±0.8 and mRMR-rf 67.2±1.6; Table 6 (COVID, k=16 and k=32) puts YOTO second behind mRMR-f. The paper selectively emphasizes F1/AUPRC while downplaying AUROC, even though AUROC is standard for imbalanced classification. The claimed 'exception at k=16' is not an isolated outlier; YOTO's advantages at larger k are often small (<2 F1 points) and negative for AUROC. This is a correctness-risk issue: if the numbers are accurate, the claims are overstatements; if the numbers are inaccurate, the evaluation is unreliable. Either way, the central claim as stated cannot be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes YOTO, an end-to-end differentiable framework for joint gene subset selection and multi-task prediction in single-cell transcriptomics. Selection is performed by learning gene importance scores and applying a hard top-k mask in the forward pass while using a relaxed permutation matrix (based on the Plackett-Luce model) in the backward pass, enabling gradient-based training of the discrete selection. The authors evaluate on two datasets (mouse VISp spatial transcriptomics and COVID-PBMC scRNA-seq), comparing against Seurat, mRMR variants, Block HSIC Lasso, and PERSIST, and report F1, accuracy, AUROC, and AUPRC across subset sizes k = 16–256. The central claim is that YOTO \"consistently outperforms state-of-the-art baselines\" while providing a single-model multi-task solution.","tokens_in":13873,"tokens_out":2398,"duration_ms":27010,"significance":"If the claims were fully supported, the contribution would be practically valuable: a single-stage, sparse, end-to-end feature selector with a built-in classifier eliminates the two-stage pipeline overhead, and the multi-task formulation allows partially labeled datasets to share information. The paper includes a well-motivated ablation isolating the selection module, reports results over three seeds, and examines multiple imbalanced-classification metrics. However, the paper's own tables do not support the \"consistently outperforms\" claim, and at several evaluated subset sizes and metrics YOTO is clearly worse than baselines. The overstated central claim, together with the lack of any diagnostic for the straight-through gradient approximation, requires revisions before the paper can be accepted.","major_comments":[{"comment":"The abstract's claim that YOTO \"consistently outperforms state-of-the-art baselines\" is contradicted by the paper's own results. On VISp at k=16 (Table 5) YOTO achieves F1 64.0±4.0 versus PERSIST 74.3±0.8 and mRMR-rf 67.2±1.6; on COVID-PBMC (Table 6) YOTO is second at k=16 (71.5 vs. mRMR-f 78.0) and at k=32 (79.5 vs. 81.7). The text acknowledges only \"one notable exception\" at small panel sizes, but these are multiple settings where YOTO trails. The claim needs to be weakened to \"competitive and often superior at moderate/larger panel sizes,\" or the experiments and analysis need to be expanded to justify the original wording.","section":"Abstract and §4.4.1, Tables 5 and 6"},{"comment":"The robustness claim across metrics is undermined by the AUROC results. At k=64 on VISp (Table 7), YOTO's AUROC is 91.9±0.1 while mRMR-rf and Block HSIC Lasso reach 99.9±0.0. On COVID-PBMC (Table 8), YOTO's AUROC is 80.9±0.2 versus 98.3±0.2 for mRMR-f and 95.8±0.1 for PERSIST — a 17-point gap. The text describes this as \"slightly lower,\" which is a mischaracterization. Since AUROC is a standard metric for imbalanced classification and the paper claims robustness across metrics (§4.4.3), the paper must either report these results honestly, provide a concrete reason why AUROC is not informative here beyond a general expectation, or revise the claim.","section":"§4.4.3, Tables 7 and 8"},{"comment":"The entire method rests on the unpublished assumption that the straight-through estimator (hard forward mask, relaxed permutation matrix in the backward pass) provides a faithful gradient signal for the discrete top-k selection. The paper provides no analysis, diagnostic, or comparison of the learned score rankings against the performance of the hard mask. For example, one could test whether the selected genes at convergence are stable across seeds or whether the relaxed gradients mis-credit non-selected genes; the k=16 underperformance may reflect this mismatch. Since the method is not derivable as a proper estimator for the discrete objective, a correctness-risk concern remains. The authors should add at least one diagnostic or a formal argument that the relaxed gradient correlates with the true discrete selection objective.","section":"§3.1, Eq. (3) and straight-through estimator"}],"minor_comments":[{"comment":"The tables and figure use inconsistent formatting: 'celltype5' in the text and 'celltype5' in Table 8 caption, but 'celltype5' in Table 6; also 'cell_types_25' appears as a task name. The method row is labeled 'Ours' in tables but the paper name is YOTO; unify terminology.","section":"Tables 5–8 and Figure 2"},{"comment":"Figure 2 omits Seurat and mRMR-f 'due to consistently low performance,' but these baselines appear in Tables 5 and 6 and are discussed. Omitting them makes the figure less complete; it would be clearer to include all baselines or state explicitly that only a subset is shown for visual clarity.","section":"§4.4.1 and Figure 2"},{"comment":"The notation π(s) is used both for the differentiable permutation matrix and as an argument in P(π(s)), which is confusing because π(s) is a matrix, not a permutation. Use different symbols, e.g., M(s) for the matrix.","section":"§3.1, Eq. (2)"},{"comment":"The ablation says 'temperature parameters' are identical, but no temperature schedule is specified. Provide the annealing schedule or a reference to where it is given, so the equivalence is verifiable.","section":"§4.4.4, Table 4"},{"comment":"Tables 7 and 8 are presented as extended results but are crucial for the AUROC discussion in §4.4.3; consider moving them to the main text or explicitly cross-referencing them earlier.","section":"Appendix A.2"}],"recommendation":"major_revision","confidential_remarks":"The central claim of consistent superiority is not supported by the paper's own data; however, the methodological idea (end-to-end sparse selection with a built-in classifier) is sound in principle and the negative results at small k are themselves informative. I believe the paper can be repaired by honestly reframing the claims, adding the missing straight-through diagnostic, and giving due weight to AUROC. I would not recommend rejection because the core method is novel and the ablations are well-designed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth taking seriously: combining differentiable top-k ranking with a straight-through estimator and multi-task heads gives you a genuinely end-to-end, hard-sparse gene selector with a built-in classifier. That is a useful recombination of known pieces, and the method section is clear about how it works. The multi-task framing is a real plus—one model for several label sets, with partially labeled data contributing to a shared representation—and the ablation against PERSIST under matched single-task settings suggests the discrete selection mechanism itself helps at moderate and larger panel sizes. The paper is readable and the experiments use real datasets, which counts for something.\n\nThe problem is the claims versus the reported numbers. The abstract says YOTO \"consistently outperforms state-of-the-art baselines,\" but the paper's own tables contradict that. At k=64 on COVID-PBMC, YOTO's AUROC is 80.9±0.2 while mRMR-f gets 98.3±0.2 and PERSIST 95.8±0.1—that is a 17-point gap, and Section 4.4.3 waves it off as \"slightly lower.\" The VISp table shows the same pattern: YOTO AUROC 91.9 versus 99.9 for Block HSIC Lasso and mRMR-rf. At k=16, YOTO loses on F1 to PERSIST and mRMR-rf on VISp, and trails mRMR-f on COVID. The prose calls k=16 \"the only exception,\" but it is not an isolated outlier; it is part of a pattern where YOTO's advantages are mostly at larger k and mostly on F1/AUPRC, not on AUROC. The paper is allowed to argue that F1 and AUPRC matter more in imbalanced settings, but it cannot describe a 17-point AUROC deficit as \"slightly lower\" and still claim consistent superiority.\n\nThe other soft spot is the straight-through estimator. The forward pass uses a hard binary mask while the backward pass uses the relaxed permutation; the paper assumes this proxy optimizes the true discrete selection, but gives no diagnostic or analysis of the mismatch. That is a legitimate gap, though not fatal—it is the kind of thing that can be addressed with a small experiment. Also missing: code, full hyperparameter schedules, and a comparison to other sparse end-to-end methods beyond PERSIST. These are all fixable.\n\nWho is this for? People working on gene-panel selection for single-cell or spatial transcriptomics who want an end-to-end alternative to two-stage pipelines. With revisions—toning down the claims, explaining the AUROC discrepancy, releasing code—this could be a solid contribution. As it stands, the empirical case is overstated. Still, I would send it to peer review; the method is principled and the flaws are in the presentation and evaluation, not in the core idea.\n\nBring it to reading group as a case study in how benchmark claims can outrun the tables.","headline":"A plausible method that overclaims: the end-to-end selection idea is sound, but the paper's own AUROC tables contradict the 'consistently outperforms' headline.","tokens_in":14438,"tokens_out":1521,"would_cite":false,"duration_ms":17371,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"YOTO jointly learns a discrete top-k gene subset and a multi-task predictor in one differentiable model, claiming consistent gains in F1 over two-stage baselines and the non-sparse gradient baseline once panels reach 32 genes and above.","keywords":["end-to-end feature selection","gene subset selection","single-cell RNA-seq","differentiable ranking","multi-task learning","straight-through estimator","Plackett-Luce","sparse selection"],"falsifier":"Train YOTO on a synthetic dataset where a known set of 32 ground-truth informative genes is planted among thousands of noise features, keeping the rest of the pipeline unchanged. If the learned subset, evaluated by the same F1 metric, is not statistically distinguishable from a random or uniform top-k subset, or if the gradient signal for a gene outside the top-k does not align with whether that gene could plausibly enter the top-k, then the straight-through feedback loop is not doing the work claimed.","tokens_in":13356,"feed_emoji":"🧬","tokens_out":5111,"duration_ms":57678,"temperature":0.7,"pith_summary":"The paper claims that sparse, end-to-end gene subset selection is not only feasible but better than two-stage pipelines: a single differentiable model that ranks genes, keeps only the top k in its forward pass, and runs multiple prediction tasks on the same selected genes attains higher F1-scores on single-cell RNA-seq datasets than classical selectors followed by a Random Forest, and than the non-sparse gradient-based baseline. The authors evaluate this on mouse visual-cortex cells and COVID-19 patient PBMCs, using panels from 16 to 256 genes. Across panel sizes of 32 genes and above, YOTO consistently matches or beats all baselines; at 64 genes it also outperforms them on the harder tasks. A controlled ablation attributes the gains to the sparse discrete selection mechanism itself, not to multi-task learning or model capacity. A sympathetic reader would care because biomarker panels that can be measured cheaply and used immediately, without retraining a downstream classifier, would lower the cost of spatial transcriptomics and targeted clinical profiling.","feed_headline":"A single trained model outscores two-stage gene selection from 32 genes up","feed_subtitle":"YOTO picks a sparse gene set and predicts in one training run, improving F1 on two scRNA-seq benchmarks while cutting retraining.","key_machinery":"YOTO's engine is a differentiable permutation matrix derived from the Plackett–Luce ranking model, which converts learnable per-gene scores into a relaxed top-k selection; a straight-through estimator then forces the forward pass to use a hard binary mask so only k genes enter the shared encoder, while gradients flow back through the relaxed matrix. A temperature schedule sharpens the ranking over training, and k itself is annealed from all genes down to the target panel size. This construction is what lets selection and prediction be optimized together with exact sparsity present at inference.","core_discovery":"YOTO is a single neural architecture that performs feature selection and prediction simultaneously. Each gene gets a learnable score; a differentiable permutation matrix derived from the Plackett–Luce ranking model turns those scores into a soft top-k selection, and a straight-through estimator makes the forward pass use a hard binary mask so that only the k selected genes reach the encoder and task heads. The prediction loss therefore backpropagates through the relaxed ranking, closing a feedback loop where the task guides selection and selection shapes representation. The central discovery claimed is that this closed loop, with exact sparsity during training, yields gene panels whose downs","pith_inferences":["A direct stress test would be to swap the Plackett–Luce permutation for independent Gumbel-Softmax top-k sampling; if performance holds, the exact ranking structure is not the active ingredient, pointing to the sparse mask itself.","The straight-through mismatch could be diagnosed by measuring the gradient alignment between the hard-mask selection and the relaxed permutation at identical weights; low alignment would imply the training signal is a noisy proxy for the deployed discrete choice.","Since the method's advantage appears only from 32 genes upward, the practical deployment window is spatial transcriptomics or targeted gene panels with at least that many measurements; the paper acknowledges the k=16 regime is not where YOTO wins.","Turning the same closed loop toward an unsupervised or semi-supervised objective, such as reconstructing the full transcriptome from the selected genes, could produce interpretable biomarker panels without requiring any labels, an extension the paper does not test."],"forward_implications":["With panels of 32 genes or more on the visual-cortex benchmark, and at 64 genes on the COVID-PBMC benchmark, YOTO's F1 is at or above every baseline tested, and the gap widens as panel size grows.","A single training run serves all tasks simultaneously (three per dataset), whereas baselines require three to six separate models for feature selection and prediction per task.","No downstream classifier is needed: because the forward pass is truly sparse, the model's own prediction heads directly evaluate the selected panel, eliminating retraining.","The ablation with identical architecture, single-task objective, and temperature schedule shows the sparse end-to-end selection module, not model size or multi-task supervision, drives the gains.","The learned gene sets are claimed to capture interactions and complementarity among genes, not just individually informative markers, which is why performance improves more steeply as more genes are allowed."],"fun_headline_variants":["One training run picks the genes and predicts—YOTO beats two-stage pipelines","YOTO: train once, select and predict together—no retraining needed","Closed-loop gene selection: YOTO's differentiable top-k beats baselines","Simultaneous gene selection and prediction in a single architecture—YOTO","Sparse gene subsets from end-to-end training—YOTO outperforms on scRNA-seq"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that gradients computed through the relaxed top-k permutation accurately improve the hard discrete choice; the paper assumes the straight-through estimate of the gradient is a reliable optimizer of the true discrete objective, without measuring the discrepancy between the relaxed and hard selections.","fun_headline_variants_meta":{"raw":{"variants":["One training run picks the genes and predicts—YOTO beats two-stage pipelines","YOTO: train once, select and predict together—no retraining needed","Closed-loop gene selection: YOTO's differentiable top-k beats baselines","Simultaneous gene selection and prediction in a single architecture—YOTO","Sparse gene subsets from end-to-end training—YOTO outperforms on scRNA-seq"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1193,"prompt_tokens":759,"completion_tokens":434,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":332}},"tokens_in":503,"tokens_out":434,"duration_ms":5189,"temperature":1.0,"reasoning_tokens":332,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T15:10:26.792253+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train YOTO on a synthetic dataset where a known set of 32 ground-truth informative genes is planted among thousands of noise features, keeping the rest of the pipeline unchanged. If the learned subset, evaluated by the same F1 metric, is not statistically distinguishable from a random or uniform top-k subset, or if the gradient signal for a gene outside the top-k does not align with whether that gene could plausibly enter the top-k, then the straight-through feedback loop is not doing the work claimed.","supporting_citations":[],"review_version":1}